[SPARK-10433] Gradient boosted trees: increasing input size in 1.4

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 1.4.1
• Fix Version/s: 1.5.0
• Component/s: MLlib
• Assignee: Unassigned
• Reporter: Sean R. Owen
• Created: 03/Sep/15 13:51
• Updated: 24/Mar/16 11:52
• Resolved: 24/Mar/16 11:52

Description

(Sorry to say I don't have any leads on a fix, but this was reported by three different people and I confirmed it at fairly close range, so think it's legitimate

This is probably best explained in the words from the mailing list thread at http://mail-archives.apache.org/mod_mbox/spark-user/201509.mbox/%3C55E84380.2000408%40gmail.com%3E . Matt Forbes says:

I am training a boosted trees model on a couple million input samples (with around 300 features) and am noticing that the input size of each stage is increasing each iteration. For each new tree, the first step seems to be building the decision tree metadata, which does a .count() on the input data, so this is the step I've been using to track the input size changing. Here is what I'm seeing:

count at DecisionTreeMetadata.scala:111 
1. Input Size / Records: 726.1 MB / 1295620 
2. Input Size / Records: 106.9 GB / 64780816 
3. Input Size / Records: 160.3 GB / 97171224 
4. Input Size / Records: 214.8 GB / 129680959 
5. Input Size / Records: 268.5 GB / 162533424 
.... 
Input Size / Records: 1912.6 GB / 1382017686 
.... 

This step goes from taking less than 10s up to 5 minutes by the 15th or so iteration. I'm not quite sure what could be causing this. I am passing a memory-only cached RDD[LabeledPoint] to GradientBoostedTrees.train

Johannes Bauer showed me a very similar problem.

Peter Rudenko offers this sketch of a reproduction:

val boostingStrategy = BoostingStrategy.defaultParams("Classification")
    boostingStrategy.setNumIterations(30)
    boostingStrategy.setLearningRate(1.0)
    boostingStrategy.treeStrategy.setMaxDepth(3)
    boostingStrategy.treeStrategy.setMaxBins(128)
    boostingStrategy.treeStrategy.setSubsamplingRate(1.0)
    boostingStrategy.treeStrategy.setMinInstancesPerNode(1)
    boostingStrategy.treeStrategy.setUseNodeIdCache(true)
    boostingStrategy.treeStrategy.setCategoricalFeaturesInfo(
      mapAsJavaMap(categoricalFeatures).asInstanceOf[java.util.Map[java.lang.Integer, java.lang.Integer]])

val model = GradientBoostedTrees.train(instances, boostingStrategy)

Issue Links

• duplicates: SPARK-6684 Add checkpointing to GradientBoostedTrees
• is duplicated by: SPARK-10616 GradientBoostedTrees stuck with 2958359 features train data
• is duplicated by: SPARK-10629 Gradient boosted trees: mapPartitions input size increasing